The magnetotelluric (MT) method has become more widely used in hydrocarbon exploration. The inversion of MT data, which can determine the electrical structure of subsurface, is a nonlinear and multimodal optimization problem. Particle swarm optimization (PSO) algorithm is a good solver for this geophysical inversion problem, whereas it has a shortage of heavy computation time. An improved parallel adaptive PSO inversion algorithm for MT data is proposed in order to decrease the computation time. The performance of the proposed algorithm was evaluated on the Dawn 4000L supercomputer using the synthetic MT data of 1D layered geo-electrical models of three and four layers. The numeric results show that the proposed algorithm can obtain the as good solution as the serious PSO inversion algorithm, and can reduce the computation time obviously when more computing nodes been employed. This result indicates that proposed improved parallel inversion algorithm can deal with the computation time problem and provide theory and technology support the MT data non-linear inversion based on PSO.